RefineNet: Iterative refinement for accurate object localization

We investigate a new strategy for improving localization accuracy of detected vehicles using a deep convolutional neural network. Specifically, we implement an iterative bounding box refinement on top of a state-of-the-art object detector. The bounding box refinement is achieved by iteratively pooling features from previous object location predictions. On KITTI vehicle detection benchmark, we achieve up to 6% improvement in average precision over the baseline results. Furthermore, the proposed refinement framework is computationally light, allowing for object detection at high run-time speeds. Our method runs at ~0.22 seconds per image on images of size 1242 × 375, making it one of the fastest detectors reported on the KITTI object detection benchmark.